AI Dev 25 X NYC Nicholas Clegg: How AWS Moved Beyond Orchestration with Strands SDK

Why It Matters

Strands SDK lets developers replace fragile, hand‑coded LLM pipelines with self‑orchestrating agents, dramatically speeding AI product development and reducing maintenance overhead for enterprises.

Summary

In this session, senior AWS engineer Nicholas Clegg explains how AWS transitioned from traditional, hard‑coded orchestration of large language model (LLM) calls to a model‑driven paradigm embodied in the open‑source Strands SDK. He frames the discussion around the limitations of early agent workflows—sequential LLM prompts chained together—which proved brittle, difficult to extend, and prone to error when new customer use cases emerged. The Strands SDK, now the backbone of AWS’s agentic products such as Kira IDE, QDeveloper CLI, and Bedrock, enables agents to decide their own execution path by looping between a model and a set of tools.

Clegg illustrates the shift with a network‑analyzer agent that originally followed a fixed workflow: scan, connect, list errors, and reply. When the server was simply turned off, the rigid pipeline failed to provide a useful answer, forcing developers to constantly refactor code for each edge case. By adopting a model‑driven approach, the agent can dynamically invoke a network‑scan tool, interpret the result, and suggest turning the server on without pre‑written branches. The Strands SDK reduces this to a few lines of Python, defining four building blocks—model, tools, system prompt, and context—allowing developers to plug in any LLM (Anthropic, OpenAI, Bedrock) and custom tools via a simple decorator.

Live demos reinforce the practicality of the framework: an eight‑line agent uses a calculator tool to compute a square root, a custom multiply function is exposed as a tool, and an OpenAI‑backed agent runs identically. Clegg also tackles “context bloat,” showing how wrapping a tool call inside a sub‑agent (e.g., a Web Researcher summarizer) isolates large outputs from the parent’s context window, cutting the parent’s token usage from 15,000 to 2,700 characters. This hierarchical, multi‑agent orchestration—sometimes called a swarm—demonstrates how Strands can scale complex tasks while preserving performance.

The broader implication is that developers can now build, iterate, and maintain AI‑driven applications with far less boilerplate and technical debt. By abstracting orchestration into the LLM’s reasoning loop and providing a lightweight SDK, AWS lowers the barrier for enterprises to deploy reliable, self‑correcting agents, accelerating time‑to‑value for AI‑enhanced services and potentially reshaping how software teams integrate generative AI into production workloads.

Original Description

Nicholas Clegg, Senior Software Engineer at AWS, discussed model-driven agents at AI Dev 25 x NYC.

Until recently, building AI agents meant wrestling with complex orchestration—state machines, rigid workflows, and endless error handling that still broke on edge cases. Strands SDK emerged from this exact struggle while developing Amazon Q and Kiro, replacing elaborate frameworks with a simple model-driven approach that lets LLMs orchestrate themselves.

Through live demos showcasing how this works in practice, attendees learned to bootstrap agents from scratch and build multi-agent architectures—walking away ready to implement production-ready model-driven agents with the Strands SDK, regardless of their cloud platform.

--------------

Join us at AI Dev 26 x San Francisco! Tickets: https://ai-dev.deeplearning.ai/